license: other
license_name: lfm1.0
license_link: LICENSE
base_model: LiquidAI/LFM2.5-24B-A2B
pipeline_tag: text-generation
library_name: transformers
tags:
- liquid
- lfm2.5
- moe
- uncensored
- heretic
- ara
- abliterated
LFM2.5-24B-A2B-Uncensored
A decensored version of LiquidAI/LFM2.5-24B-A2B, made with
Heretic's Arbitrary-Rank Ablation (ARA) and a distributed
Optuna search across three machines.
BF16 weights in the official layout. Only one tensor differs from the original:model.layers.38.self_attn.out_proj.weight. Every other tensor, the tokenizer and the chat
template are the official ones.
Results
Measured with Heretic's own evaluator: 100 harmful prompts (mlabonne/harmful_behaviors) for
refusals (keyword match), 100 harmless prompts (mlabonne/harmless_alpaca) for first-token KL.
Thinking is closed with an empty <think></think> block before measuring (Heretic's
Chain-of-Thought handling), so both numbers describe the actual answer.
| Refusals | KL divergence | |
|---|---|---|
| Original | 89/100 | 0 |
| This model | 0/100 | 0.260 |
Measured on these BF16 weights. (During the search, on the FP8-experts working copy, the same edit scored 0/100 at KL 0.186.)
A note on comparing KL numbers
LFM2.5 always thinks before answering. If KL is measured without closing the thinking block,
the first token is the start of the reasoning, which barely changes, and KL comes out
dramatically lower. Example: another public LFM2.5 Heretic variant measures 0.0033 that way, but
0.229 with thinking closed (same weights, same prompts). Compare KL values only when they were
measured the same way.
Method
- Heretic upstream (ARA modifier), LoRA rank 50, L-BFGS.
- Searched layers 20–39; the optimizer ended up editing only layer 38's attention output.
- Search used a copy of the model with only the fused MoE experts in FP8 (which ARA never
touches) for speed. The edit was then merged into the original BF16 weights.
Usage
Same as the original model (transformers / vLLM / SGLang / llama.cpp). Liquid AI recommendstemperature 0.1, top_k 50, repetition_penalty 1.05. The model still thinks before
answering. The official DSpark draft model
targets the original weights; acceptance on this model has not been measured yet.
⚠️ Notice
This model has had its refusal behavior substantially reduced. It may produce harmful,
inaccurate or offensive content. Intended for research, red-teaming and personal use; you are
responsible for how you use it and for complying with applicable laws and the base model license.
License & credits
Derivative of LiquidAI/LFM2.5-24B-A2B, released
under the LFM Open License v1.0 (see LICENSE). All rights to the base model remain with
Liquid AI. Decensoring by Heretic (p-e-w, AGPL-3.0) and its
ARA method.